Does the average size of a contribution matter?

Some campaigns rely on lots of donors who give small amounts, while other campaigns rely on fewer donors who give large amounts. Therefore, the average size of a contribution might say something about the types of people supporting a campaign, which might indicate a campaign is more or less likely to succeed.

I began by making a boxplot to look at the average contribution size for each campaign. I excluded negative values, which represent a campaign returning money (e.g. because the campaign ended, or because a contributor more than is allowed.)

## Warning: Removed 2746 rows containing non-finite values (stat_boxplot).

It was hard to tell much from the boxplot. The median contribution for the most successful candidates – Trump, Clinton, Sanders, and Cruz – was indeed lower than most other candidates. But that didn’t help distinguish between Trump, Clinton, Sanders, and Kruz.

If we zoomed in on the most successful candidates, would we start to see that campaign contributions did predict success?

Does the number of contributions or total amount of contributions matter?

When we started, I noted that news coverage about national trends in campaign contributions made them seem like a lousy predictor of a candidate’s success in 2016. If we focus on only Pennsylvania, will we find that campaign contributions are a better predictor at the state level?

So, when we look at Pennsylvania state trends in campaign contributions, do campaign contributions seem like a better predictor? Not really.

In Pennsylvania, Sanders received about 50% more contributions than Clinton, and he still lost the primary.

Clinton received several times the number and sum of contributions as Trump, and she still lost the general election.

If the counts and sums of contributions don’t tell us much, can we learn anything from the timing of those contributions?

Does the timing of contributions matter?

Before delving into the data, I wanted to refresh myself on the relevant dates and get a baseline for all contributions.

Source: website of primary dates and just google results pages for other dates

Contributions to the primary didn’t pick up until early 2016, then dipped around May and June, 2016 – immediately following the Pennsylvania primary. Then they jupmed again in July, when both major parties held their conventions.

After the conventions, contributions switched to the general election. These peaked in early November, near election day.

There’s nothing too surpring here. However, it’s important to note that several “primary” contributions actually occured after the primary elections, meaning they couldn’t have been useful in predicting the primary results.

So now we’ve got a baseline. But what does looking at contributions to individual candidates tell us about their odds of success?

This graph already says a lot, but I want to hone in on the primary and general election, so I’m going to limit the time frame.

So does the timing of contributions seem like it could be a predictor of success for the primaries? No.

If anything, the timing makes the number of contributions look like an even worse predictor of success, since Clinton and Trump received many “primary” contributions in the lead-up to their party conventions, after they were already the clear winners. If we ignored those late-stage contributions, Sanders’ lead over Clinton and Cruz’s lead over Trump would be even more pronounced.

Trump’s contributions didn’t even climb noticeably above zero per week until after the Pennsylvania primary.

I thought that plotting the relative number of contributions would make these trends even more noticeable, by eliminating overall trends in contributions.

Looking at the relative numbers, it’s clear that Sanders received about twice as many contributions per week as Clinton from late 2015 to May 2016. But, again, those contributions didn’t mean he would win.

It’s possible that the timing of contributions might still be important, but that we should be looking at contribution amounts. For example, if a candidate received a greater amount of contributions closer to the election, maybe that indicated support for their campaign was growing. What do the relative amounts of contributions tell us about the growing or waning support for a campaign?

When we look at the timing of when contribution amounts were received, do we see a different story? No – Sanders received a greater amount of contributions than Clinton in the months immediately before the primary. The same is true for Cruz versus Trump.

There’s another possible interpretation of timing. Clinton received a higher amount of contributions for most of November 2015 through February 2016. We might tentatively posit that receiving greater amounts when Clinton did – before primaries began – was a predictor of later success.

But that doesn’t hold true on the Republican side, where Cruz, Carson, and Rubio all led Trump for late 2015 and early 2016.

Is it possible that the timing of contributions isn’t a good predictor during the primaries, but that it’s a better predictor of the general election?

Clinton absolutely dominated Trump in the number of contributions received per week from mid-August to the election day. It’d be difficult for there to be a clearer indication that the number of contributions did not predict the winner.

As for the amount of contributions?

The gap in contribution amounts was consistently closer, but Clinton still received more than twice – and sometimes three times – the amount of contributions as Trump.

So it didn’t matter when a candidate received their contributions. Did it matter who was giving the contributions?

Do the occupations of contributors matter?

Perhaps some contributors more reliably indicate the eventual winner. Some people might be more influential in an election’s outcome, or a certain group of people might be a good bellweather for how their neighbors will vote.

I didn’t have a detailed profile on each contributor, but I did have information about their occupation and their geographic location. I decided to first look at contributors’ occupations.

Which candidate did better with nurses? With CEO’s? And were contributions from one group possibly more predictive than support from another group?

To answer these questions, I wanted to look at the number of unique contributors with each occupation. If John, a truck driver, contributed four times, I still only wanted to count him as one truck driver.

To identify unique contributors, I decided that two contributions that shared the same contributor name, zip code, and occupation were probably from the same person.

## [1] 61381

Woah! There are only 61,381 unique contributors. That’s far less than I expected, given that there were more than 241,000 positive-value contributions. This is worth investigating more!

But first, I still wanted to look at occupations.

There’s a lot to unpack here.

For one, Kasich’s main supporters were executives, presidents, and CEOs – though since he didn’t lead among these occupations, Kasich’s loss doesn’t say anything about the influence of these wealthy contributors.

There were two occupations that are difficult to interpret – why did some people list their occupation as “None,” and why were these mostly Cruz contributors? And why did Trump receive the most contributors for whom their occupation was “Information Requested”? It’s possible these people were unemployed, but it’s also possible these people chose not to disclose their occupation, which begs the question Why? The dataset doesn’t have enough information to answer this question.

As for Sanders, he fared well among contributors of several occupations, many of which also favored Clinton. There were a few exceptions: Sanders and Trump split the support of engineers, and Sanders led all the candidates among software engineers and the not employed.

Compared to Sanders, Clinton had an uncontested lead among several more common occupations. This is really the first time that Pennsylvania’s campaign contributions begin to display some predictive power. But is that because we’re looking at occupations, or because we’re looking at unique contributors?

The same question arises when we look at Trump — who either led or held even with the other Republicans in almost every category.

But before looking at unique contributors more generally, let’s look at unique contributors to the general election.

When shifting focus from the primaries to the general election, we can ask an interesting question: Did Sanders’ supporters switch to Clinton or Trump? Among occupations where Sanders clearly led in the primary — software engineers and the not employed — Clinton now leads Trump. These results are suggestive, but certainly not definitive, since we’re not necessarily tracking the same individuals.

Then there’s the larger question: Could focusing on contributors’ occupations have helped us predict the general election outcome? It’s hard to say, but we can rule out some things about occupations being predictive:

Clinton led among several occupations that might be called intellectuals or well-educated – such as attorneys, scientists, students, professors, nurses, physicians, teachers, educators, software engineers, and writers – and the support of the well-educated did not translate into a win.

The not employed and difficult-to-understand “None” category also overwhelmingly supported Clinton — and these groups’ support did not translate into electoral success.

It becomes harder to make even tentative conclusions when looking at the occupations of Trump’s supporters. Trump led among some blue collar workers — like truck drivers and contractors — but we can’t say these supporters predicted his success, only that we can’t rule it out.

But let’s now turn to a more promising question. Since about…

## [1] 0.7453366

75% of contributions were made by repeat contributors, what happens when we begin looking at the number of unique contributors, not the number of contributions?

Does the number of unique contributors matter?

Looking at unique contributors, we begin to see the possibility that campaign contributions might be predictive of a candidate’s success!

Clinton had more unique contributors than Sanders in the primary, and she won PA’s democratic primary. Trump had more unique contributors than Kasich or Cruz in the primary, and he won PA’s Republican primary. Unique contributors looks like it might be predictive of who’s going to win a party’s primary.

But the predictive power falls away when we look at the general election. Although Clinton’s lead has narrowed — as compared to contribution numbers or total dollar amounts — she still had more unique contributors in the general election and lost to Trump.

But what if we look at the timeline of when unique contributors first gave to a campaign? Will we find a way to fine tune our analysis?

Now that we’re looking at unique contributors, we get a VERY different perspective on campaign contributions during the presidential campaign.

Trump had two huge spikes in the number of first-time contributors in July, before the Republican National Convention, and in August, after the convention. Maybe when considering whether first-time contributors might predict the general election winner, we should look at unique contributors during both the primary and general election.

Also of note, Cruz was beating Trump in first-time contributors until the beginning of May, right after the Pennsylvania primary. If first-time contributors were as predictive of a primary’s winner, shouldn’t we focus on contributions before the primary took place? Otherwise, we’re just seeing contributors react to the winner.

What happens when we incorporate these lessons into the bar graphs above?

Now, it seems like our most recent — and tentative — hypotheses have been turned upside down.

The number of unique contributors was NOT predictive of a campaign’s success in the primary. Before the PA primary, Sanders had more unique contributors than Clinton, and he still lost the primary. Before the PA primary, Cruz had many times more unique contributors than Trump — and even Kasich had about twice as many contributors as Trump — and both Cruz and Kasich lost.

On the other hand, the number of unique contribtors may possibly have been predictive of the general election results. Trump had more unique contributors than Clinton throughout the entire election season.

This is the first time that some portion of campaign contribution data might be useful in predicting general election results! Certainly, this is tentative, but it’s still a potentially useful result that’s worth exploring more.

Now let’s look at one more variable: location. Much has been made of the divide between rural and city voters. What happens when we look at these voters using our newfound awareness of the potential predictive power of unique contributors?

Do the locations of contributors matter?

Before answering this question, I needed to do a little data cleaning. When I used zip codes from the campaign contributions to look up latitude, longitude, city, and state; I was surprised to find that some zip codes were not in Pennsylvania.

## [1] 27

Since there were only 27 unique contributors outside Pennsylvania, out of about 61,000 contributors, I decided to simply exclude these contributors and continue with the analysis.

To map contributions, I used the choroplethr package. It was mostly straight-forward, but this stack exchange post was particularly useful in helping understand choroplethr’s zipcode functionality.

A couple things to note about the above choropleths: (1) zip codes with no contributors are black and (2) some zip codes have values above 45 — however, those zip codes have the same color scheme as counties with 45 contributors

Now that that’s out of the way, what do the above maps tell us about rural Pennsylvanians versus city Pennsylvanians?

As for rural areas, Trump certainly had more contributors than Clinton. However, Clinton had a small number of contributors in many rural zip codes.

As for cities, especially Philadelphia, the above maps don’t helps us distinguish between the candidates. Because there were so many more contributors in urban zip codes, in order to see differences in rural areas, I needed to limit the max value on the scale. What can we learn about contributors in Philadelphia by focusing on zip codes in Philadelphia and changing the scales?

Looking at unique contributors in Philadelphia, we find the converse of rural areas. Clinton certainly had more contributors than Trump, though Trump had a small number of contributors throughout most of the city.

The numbers of unique contributors in different parts of a state seems like a possible predictor of how voters in those areas will vote.